Nobody got AI-fluent from a video course. Fluency is a rep count.

Finish a good AI tutorial and you feel sharper. You recognize the techniques, you could explain them to a colleague. Then Monday arrives, a real deadline sits in front of you, and you do the work the old way. Nothing failed. You just built the wrong thing: recognition, which is knowing a move when you see it, instead of capability, which is producing the move under pressure when nobody is showing you.

Every skill has this gap. You can watch a hundred hours of cooking shows and still panic at a hot pan. AI is no different, except the marketing around it pretends otherwise. The prompt that looked obvious in the video falls apart on your messy spreadsheet, your half-written draft, your client's weird formatting. That's not evidence you're bad at AI. It's evidence that fluency forms on your real work or it doesn't form at all.

So the question of how to get good at AI has a boring, reliable answer: reps on a task you already do every week. Waiting for more content, a better course, or a smarter model just postpones them. Here's the formation pattern we see over and over in people who cross from dabbling to fluent.

There's a serious objection to everything above, and it deserves a direct answer. The cognitive-offloading research is real, and it isn't flattering. Michael Gerlich's 2025 study of 666 people found a strong negative correlation between heavy AI use and critical thinking, mediated by cognitive offloading: the more reasoning you hand to the machine, the quieter your own gets. The same year, an MIT Media Lab team put the phrase 'cognitive debt' into circulation with an EEG study of essay writers. The ChatGPT group showed the weakest brain connectivity of the three groups tested, and many couldn't accurately quote their own essays minutes after writing them. If getting good at AI were a raw rep count, those participants should have been getting sharper. They were checking out.

But look at the rep those studies measured: paste the task, accept the output, move on. Delegation with the judgment switched off. Gerlich's own data points at the way out. The participants who held their critical thinking scores despite heavy use were the higher-educated ones, the group he found most likely to cross-check AI output against other sources instead of accepting it, and he's careful to say the correlation could run either way: weaker critical thinkers may simply lean on AI more. The rep that builds fluency has three parts, brief, review, correct: you specify the work, you judge what comes back, you fix the gap. The first part is delegation. The other two are what Anthropic's AI fluency framework calls discernment, and an AI-first rep exercises them harder than unaided work does. You aren't offloading the work. You're moving your effort from production to direction and quality control, which is a different skill and a harder one.

The opposite camp says skip the structure entirely. Ethan Mollick's standing advice is to just use AI on things you do for work or fun for about ten hours, and you'll figure out a remarkable amount. He's right, as orientation. Ten hours of poking builds intuition for the jagged frontier, the strange map of what models handle well and what they fumble. What osmosis doesn't build is the habit. BCG's AI at Work 2025 survey of 10,635 employees across 11 countries found that 79% of people who got more than five hours of training, with coaching and live sessions in the mix, became regular AI users, against 67% of those who got less. Only 36% said their training was enough. Exposure starts the engine. Reps with feedback keep it running on a Tuesday deadline.

There's a failure mode on the far side of the first win. You get one workflow humming, it feels great, and six months later you're still doing that one workflow at that same difficulty. Comfortable, useful, and stalled. Fluency plateaus when every session is the same task at the same level, the way a runner plateaus running the same easy loop every day.

Progression means handing AI more of the judgment-adjacent edges of the work; extra volume alone won't do it. If AI drafts your report, let it propose the structure next time. If it summarizes the research, ask it to flag what's missing. Each handoff feels slightly uncomfortable, which is the point: the discomfort is the rep. You stay the editor and the judge, but the territory you delegate keeps growing, and so does the skill.

If you want to know what the moment of takeoff feels like from the inside, when reaching for AI stops taking willpower, I wrote about that in the AI adoption inflection point. And if you've watched a whole company fail at this, the same recognition-vs-capability gap sits at the heart of why most corporate AI training fails. The organizational disease is the individual one, bought in bulk.

The individual cure is cheap. One workflow, three honest reps, feedback on technique, then widen. If you'd rather not run the loop alone, Candova pairs you with Cando to coach every rep on your real work. Either way, stop collecting courses. Start counting reps.